How large language models can help prevent the next pandemic

The COVID-19 pandemic was a calamity in more ways than one: lost lives, shuttered businesses, closed borders, broken supply chains, joblessness, and frayed social unity. It’s no wonder that hundreds of articles have already been written about “how to prevent the next pandemic”. Individuals, entrepreneurs, organisations, and governments all want to be able to anticipate what’s coming next and take action to avert its worst impacts.

I believe the structures currently in place for detecting and responding to outbreaks are unequal to this task. Their most important actors, governments and public health agencies, are prudent by nature and work at a deliberate pace to understand how outbreaks have spread. They are also working with limited resources and budgets that, from my observations, are geared towards more routine public health responsibilities, not the intensive and time-sensitive needs of new and emerging infectious diseases. As a result, by the time any official declaration is made, both the disease and the misinformation around it could have already travelled the world.

The benefits of predictive intelligence: LLMs

And yet, digital technology can enable reliable, actionable information to travel just as fast. Everywhere in the world, local publications, broadcasters, and digital sources provide information on events and outbreaks as they occur. The challenge lies in sifting through all those primary sources, reported in hundreds of different languages, to detect reliable patterns amid the mountains of raw, unstructured information. I have often had to analyse such sources myself as part of my work, building spreadsheets and drawing diagrams to organise available information into timelines and figure out what is happening in a remote corner of the world. This is a process that can take days. And I know that this effort is routinely duplicated by many researchers and analysts across the world.

Instead of manual work, I believe systems could be put in place to help detect early signs of disease spread, assess the disease, and provide actionable insight - what I call “predictive intelligence”. Artificial intelligence tools such as large language models, the technology that powers applications such as ChatGPT, can provide some of the support needed to make such systems possible.

LLMs combine the capability to search vast amounts of data with AI-driven learning, which allows them to contextualise and summarise information. Beyond broad public applications such as Chat GPT, LLMs can be developed for a specific purpose, instructed on which types of information patterns are most relevant, and refined to filter out inappropriate and irrelevant data.

LLMs have been widely touted for their potential application to a wide range of business functions, such as increased sales efficiency, hyper-personalised customer service, and employee training. Based upon my work as an infectious disease physician, I believed that LLMs could facilitate the painstaking work needed to identify emerging outbreaks. That's why my firm, BlueDot, is currently integrating our own LLM into our core function: an outbreak intelligence platform that can support business resilience in the face of future pandemics.

Retaining the human element to go beyond human capabilities

Anyone who has experimented with a ChatGPT prompt has seen LLM capabilities in action, including their occasional unreliability and susceptibility to misinformation. LLMs are only as precise and discerning as their human programmers make them. Yet, that's where their true potential lies. An LLM can be trained by curating its source data and supervising machine learning with human expertise. When it comes to disease surveillance, this is no small task: it requires oversight by epidemiologists, immunologists, virologists, environmental scientists, and more.

But the result is worth the effort. A purpose-built, properly trained LLM can do all the primary-source work of disease surveillance in a matter of seconds to a depth of detail no human can match. It can sift through millions of sources daily to find reports of specific diseases and syndromes. It can isolate locations, assemble timelines of events, and cross-reference sources. They can also help project the course of a disease and disseminate information quickly in the form of actionable reports tailored to the needs of a specific business.

In the current system, governments and public health agencies are the dominant response actors, but with predictive intelligence powered by LLMs, everyone, and every organisation, becomes a response actor. Businesses can move faster to protect their employees, their customers, and their markets.

Any business with a global supply network can use predictive intelligence to shore up supply chains before disruptions occur, sharing their reports with partners and providers and acting in concert to safeguard deliveries. Retailers can act faster to secure PPE for employees ahead of potential supply shortages. In the travel sector, airports and airlines can anticipate changes in travel demand and adjust routes accordingly. Ground transporters can take proactive measures to avoid cross-border slowdowns. Medical-industry firms can anticipate demand spikes for key products and devices. Pharmaceutical firms can increase production of existing vaccines and therapies and get a head start on new ones. And firms across all industries can take steps to mitigate the problems caused by employee absenteeism, whether by mandating the use of PPE, intensifying hiring and training, or reorganising shifts and schedules.

Speed and agility for future pandemic response

In the wake of COVID-19, most businesses recognise the need to have policies and plans at the ready for managing through a future pandemic. But the most important factor in pandemic response is speed and agility. LLMs can provide all businesses with the predictive intelligence they need to act earlier and more decisively than ever in the face of pandemic risks, and to adjust their response as circumstances evolve.

Imagine if, during the COVID-19 pandemic, the world had this kind of LLM-powered predictive intelligence at its disposal, and that firms responded quickly and simply to protect their business interests. We would likely have been able to slow the spread of the disease and mitigate against the pandemic's worst economic impacts. The shorter the delay between information and action, and the greater the number of actors, the better the chances that we can avert the worst impacts of any future pandemic.